Research map / Large language models / Language, safety and interpretability
Enhancing Entertainment Translation for Indian Languages Using Adaptive Context, Style and LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CASAT, a training-free framework that estimates session context (via RAG over LLM-generated plot summaries) and style (via a domain adaptation module of POS/syllable/punctuation/emotion features) to build prompts guiding LLMs for English-to-Indic entertainment translation.
Problem
Sentence-level NMT for entertainment dialogue ignores context, mood, style, and cultural nuance, and existing prompt-based or translation-memory methods are static and do not adapt to time-varying scene context in dubbing/subtitling.
Approach
Segment dialogue into genre-consistent sessions, retrieve plot context from a vector database of LLM-generated plot summaries via advanced RAG, and combine it with a session/dialogue-level style prompt (content/function words, syllabic words, modal words, idioms, punctuation-based intent, LLM-extracted emotion) into a time-varying prompt for any LLM.
Key results
- CASAT improves COMET over base prompts for all 7 LLMs across En-Hi, En-Ben, En-Tel; e.g., Llama3 70B En-Hi COMET rises from 0.63 to 0.70 (BLEU 7.96 to 9.84).
- CASAT-augmented Gemma2 27B achieves win-ratios of 58-65% vs IndicTrans2 and 61-68% vs NLLB in some directions (e.g., CG27 vs NLLB 68% En-Tel).
- Ablations show context-only and DAM-only each help, with combined CASAT giving best COMET and win-ratio across LLMs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CASAT context and style prompts | Mistral 7B · En-Hi | COMET: 0.41 -> 0.42 | Helped |
| Add CASAT context and style prompts | Mistral 7B · En-Ben | COMET: 0.43 -> 0.48 | Helped |
| Add CASAT context and style prompts | Mistral 7B · En-Tel | COMET: 0.42 -> 0.41 | Hurt |
| Add CASAT context and style prompts | LLaMa-3 8B · En-Hi | COMET: 0.51 -> 0.56 | Helped |
| Add CASAT context and style prompts | Gemma2 9B · En-Ben | COMET: 0.68 -> 0.75 | Helped |
| Add CASAT context and style prompts | Gemma2 27B · En-Ben | COMET: 0.70 -> 0.77 | Helped |
| Add CASAT context and style prompts | Aya23 35B · En-Tel | COMET: 0.44 -> 0.48 | Helped |
| Add CASAT context and style prompts | LLaMa3 70B · En-Ben | COMET: 0.66 -> 0.75 | Helped |
| Add CASAT context and style prompts | GPT-3.5 Turbo · En-Tel | COMET: 0.66 -> 0.78 | Helped |
| Add context retrieval without DAM (ablation) | Mistral 7B · En-Hi | COMET: 0.41 -> 0.41 | No clear effect |
| Add DAM without context retrieval (ablation) | Mistral 7B · En-Hi | COMET: 0.41 -> 0.41 | No clear effect |
| Add context retrieval without DAM (ablation) | Llama3 8B · En-Hi | COMET: 0.50 -> 0.55 | Helped |
Surprising findings
- Gemma2 9B and 27B match or beat Aya23 35B and Llama3 70B on COMET for En-Ben and En-Tel.
- Style-only prompting has minimal effect on translation quality, while context-only prompting helps more.
Limitations
- Relies on offline session segmentation and offline plot generation, not fully online.
- Evaluation datasets are web-scraped subtitles without public release; small test sizes (800 samples per language).
- Win-ratio judged only by GPT-4o; segmentation uses coarse three-genre classes and needs speaker/diarization improvements.
Experimental setup
Datasets: OpenSubtitles movie subtitles (Adipurush, Pushpa, Interstellar, Wolves, Maharaja, Without Remorse, Bumblebee) children's cartoon series episodes
Benchmarks: OpenSubtitles-derived En-Hi, En-Ben, En-Tel test sets
Metrics: SacreBLEU COMET (wmt22-cometkiwi-da) GPT-4o win-ratio
Baselines: Base LLM with creative prompt Context Only ablation DAM Only ablation IndicTrans2 NLLB Surrounding-sentence context
Builds on: retrieval-augmented generation in-context learning style transfer domain adaptation module
Closest papers by meaning
- From Utterance to Vividity: Training Expressive Subtitle Translation LLM via Adaptive Local Preference Optimization (ICLR 2026 · similarity 0.87)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.85)
- Calibrating Translation Decoding with Quality Estimation on LLMs (NeurIPS 2025 · similarity 0.83)
- Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass (ICLR 2025 · similarity 0.83)
- Teaching Large Language Models to Translate with Comparison (AAAI 2024 · similarity 0.82)
- Adaptive Few-shot Prompting for Machine Translation with Pre-trained Language Models (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.